LAMaS slashes multi-agent system latency by 50% without accuracy loss
New framework cuts inference latency in half while keeping accuracy high.
Multi-agent systems (MAS) coordinate multiple LLM-powered agents through structured workflows, but multi-step execution and repeated model invocations cause high inference latency. Existing orchestration methods optimize task performance and inference cost, leaving latency largely unaddressed. The end-to-end latency is governed by the critical execution path, so reducing total cost alone does not reliably reduce latency. Moreover, naive latency optimization can misassign operator-level credit and degrade accuracy.
To solve this, Xi Shi, Mengxin Zheng, and Qian Lou propose LAMaS (Latency-Aware Multi-agent System). At training time, it learns latency-aware execution graphs through constrained optimization with critical-path-aware credit assignment. At inference time, a lightweight controller adaptively eliminates redundant future agent interactions as execution unfolds. Evaluated on four benchmarks, LAMaS reduces end-to-end latency by over 50% while maintaining competitive or better accuracy. The framework is modular and transfers to other multi-agent systems with minimal changes, consistently yielding latency reductions.
- Addresses the critical-path latency bottleneck in LLM-powered multi-agent system workflows.
- Uses critical-path-aware credit assignment during training to optimize latency without accuracy loss.
- Achieves over 50% latency reduction across four benchmarks and transfers to other MAS with minimal changes.
Why It Matters
Faster multi-agent coordination enables real-time AI workflows for complex reasoning tasks.